Research map / Reinforcement learning / Policy gradient and actor-critic
Importance Resampling for Off-policy Prediction
Matthew Schlegel, Wesley Chung, Daniel Graves, Jian Qian, Martha White
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Resample replay-buffer transitions in proportion to importance ratios and apply on-policy updates instead of reweighting updates | Off-policy prediction · Four Rooms | convergence rate by number of updates | Improved |
| Resample replay-buffer transitions in proportion to importance ratios and apply on-policy updates instead of reweighting updates | Off-policy prediction · Continuous Four Rooms | convergence rate by number of updates | Mixed |
| Resample replay-buffer transitions in proportion to importance ratios and apply on-policy updates instead of reweighting updates | Off-policy prediction · Random Walk Markov Chain | update variance | Improved |
| Resample replay-buffer transitions in proportion to importance ratios and apply on-policy updates instead of reweighting updates | Off-policy prediction · Random Walk Markov Chain | learning-rate sensitivity | Improved |
| Resample replay-buffer transitions in proportion to importance ratios and apply on-policy updates instead of reweighting updates | Off-policy value prediction · TORCs | mean squared return error | Improved |
| Add bias correction by multiplying resampled updates by the buffer's average importance ratio (ablation) | Importance Resampling · TORCs | mean squared return error | Hurt |
| Use IR+WIS-TD(0) instead of IR with a constant learning rate (ablation) | Incremental off-policy prediction · Continuous Four Rooms | convergence rate by number of updates | Improved |
| Increase the aggressiveness of V-trace importance-ratio clipping (ablation) | V-trace · Random Walk Markov Chain | variance | Mixed |
| Use WIS-Minibatch or V-trace instead of IS (ablation) | Off-policy prediction · Four Rooms | learning performance | Improved |
| Use WIS-Buffer instead of IS (ablation) | Off-policy prediction · Four Rooms | learning performance | No clear effect |
Closest papers by meaning
- Importance Sampling Techniques for Policy Optimization (JMLR 2020 · similarity 0.90)
- Towards Optimal Off-Policy Evaluation for Reinforcement Learning with Marginalized Importance Sampling (NeurIPS 2019 · similarity 0.90)
- Subgaussian and Differentiable Importance Sampling for Off-Policy Evaluation and Learning (NeurIPS 2021 · similarity 0.90)
- Demystifying the Paradox of Importance Sampling with an Estimated History-Dependent Behavior Policy in Off-Policy Evaluation (ICML 2025 · similarity 0.89)
- Scaling Marginalized Importance Sampling to High-Dimensional State-Spaces via State Abstraction (AAAI 2023 · similarity 0.89)
- Kernel Metric Learning for In-Sample Off-Policy Evaluation of Deterministic RL Policies (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
